BillyBag2's profile - activity

That's a good idea, this makes it self priming. I still want to know the orientation of the car as it changes. TLD will self learn as the car rotates, but will not feed this back to me an absolute orientation. I think I still need to take images of the target and detect the orientation. But now I'm detecting in a small scene. Also I do not need to know orientation as quickly as position, as position will change more quickly than the orientation. This buys me time to do the eight detects, unless you guys can explain to me how to get orientation in a single pass.

Looking at https://play.google.com/store/apps/de... TLD seams ideal for doing tracking of an object using a hand held camera. It does not solve the issue of which way round it is. Perhaps use TLD to detect a car in a scene by the user selecting it, but I then have to work out which way around it is, probably using a pre trained detector.

For this problem there will only be zero or one cars in the scene, it will always be the same car, just different directions . I want to use hand held camera and also detect the car when stationary, so movement type detections will be very problematic.

The car is rotating in a different axis to the camera. Only when looking straight down at the roof of the car is the image rotation and car rotation the same. Looking horizontally I may see the back of the car or the front. I need a little rotation invariance for when the camera is not held level, but that is a side issue. Unfortunately the car is not cuboidish. Its open wheel. So spotting the warped "faces" of the car is not an option, in this case. (Sorry I do not have a set of training image to post, yet. I wiil try and do so in the next few days.)

I think the bit I missed was the relationship between field of view and focal length. This looks quite easy too. See http://paulbourke.net/miscellaneous/l....

The difficult question is do you have the effective focal length or the actual focal length? Probably the effective focal length. I which case the width is 35mm. If its the true focal length (which I don't think you have enough information to know) use the sensor size from the EXIF data.

So you fov is probably, 2 x aTan( 0.5 x 35mm / (effective focal length in mm))

Not sure of the relationship between the parameters you have and the field of view. However if you can get this I think an approximate calculation of distance would be easier.

Angle of arc = (field of view) * (object pixel width)/(screen pixel width).

Then tan(arc/2) = (true width/2)/(true distance).

So (true distance) = width/2.tan(arc/2). (very approximate.)

May contain mistakes, lots of simple assumptions made, low tech answer.

(PS lots of info in the exif data from a still from the camera. Something may be useful. I think its an augmented reality trick to get your app to take a sample picture and read this info to get approximate camera geometry.)

I want to detect only one of N objects. The objects are similar. I can train for each object including the other objects as negative images, along with other negative images.

I can then detect each object. I should only get one hit. I can pick the one with the best match if more than one hit.

My question is, is there a better way to do this than running a detection for each object?

Can I detect the one from N objects is a single go, ether using all the training results or by training in a different way?

(I'm happy to create a new tool using OpenCV API if required.)

EDIT

One project in mind to detect a model car and its crude orientation. It will be on a flat surface with the camera "level" but the camera could be at any elevation. To make the task easier for feature based stuff I choose (rightly or wrongly) eight positions where the features would be different. Front of the car, 45 degrees ie front top and one side. 90 degrees, side no front or back, etc etc. The big task is collecting training material so 'm trying to plan ahead a bit.

Steven's plan looks like a better one where I build a car detector and then a "found car" orientation detector. For the car detector was hoping to use Haar features and standard tools to create a car detector. (Maybe LBP for performances issues.)

Its very unclear to me what data I would feed the second stage (SVM?)

• I could get the cropped image of the found car
• I could possibly feed it the original found features from the whole image cropped to the car region. Saving the feature finding step again.
• Possibly just the features car finder used to find the car. Already filtering irrelevant features.

EDIT2

Here are some images from the internet of the sort of thing...

Above is what I would call a south east facing car.

Above is a south west facing car.

I would like to

• Detect the car in a scene.
• Detect which way it is facing (one of eight)

I want to avoid creating eight different "car in scene" detectors if this is not a cleaver way of doing things. (If creating eight detects is the "best" I can do that.) (These are not training pictures just examples. Training pictures would be the car in a real scene, unfortunately I don't have a huge infinity curve for the car to drive on. )